CN105205500A - Vehicle detecting method based on multi-target tracking and cascade classifier combination - Google Patents

Abstract

The invention discloses a vehicle detecting method based on multi-target tracking and cascade classifier combination. The method specifically comprises the steps of extracting and screening 15 Haar-like features of a sample, and setting classifier training parameters for classifier training; conducting new target detection based on multi-scale filtering and conducting target tracking. The vehicle detecting method has the advantages of being high in detection precision and high in processing speed.

Description

Based on the vehicle detection that multiple target tracking and cascade classifier merge

Technical field

The present invention relates to image procossing and computer vision field, particularly, relate to a kind of vehicle checking method merged based on multiple target tracking and cascade classifier.

Background technology

Flourish in recent years along with China's communications and transportation cause, the investigation and application of intelligent transportation system (ITS) is more and more paid attention to.Accurately, real-time, complete traffic information collection is the basis of ITS, and wagon detector is then the infrastructure of dynamic information being carried out to Real-time Collection.Along with the develop rapidly of software and hardware technology, various types of wagon detector emerges rapidly.Mainly contain inductive coil detecting device, magnetic force detecting device, microwave detector, ultrasonic detector, infrared detector and video detector etc., in current China preventing road monitoring system, using maximum is inductive coil wagon detector, video-based vehicle detection and microwave vehicle detecting device 3 kinds.

Inductive coil detecting device is earth-buried detecting device, directly can provide that vehicle occurs, vehicle is by telecommunication flow informations such as, vehicle count and lane occupancy ratios.Ordinary Rd all can be arranged, be mainly used in before and after charge station, interchange, tunnel section, urban road, the occasion such as parking lot.The early investment of inductive coil detecting device is less, reliability is high, but safeguards, reinstalls difficulty, and need close track, destroy road surface, operation cost is higher on long terms, this be also restriction its continue a fast-developing lethal factor.

Microwave detector is a kind of radar detedtor being operated in microwave frequency band, and the non-buried detecting device of the traffic flow essential informations such as energy inspection vehicle flow, speed, lane occupancy ratio and vehicle, centre frequency is 10.525GHz, and working method is active.Under weather extremes, performance is outstanding, can all weather operations; Side direction mode can detect multilane; Static vehicle can be detected; Direct-detection speed.But when road has the dividing strip of irony, or accuracy of detection declines when trackside has a barrier; Detecting device mounting condition requires higher, needs rearmounted distance when side direction is installed; Rate accuracy is low.

Computer Vision and computer graphical recognition technology combine by video encoder server technology noveldata acquisition technology, development in recent years is rapid, represents the developing direction of future transportation stream information detection field.It is as sensor with video camera, in range of video, virtual coil is set, i.e. detection zone, vehicle makes background gray levels change when entering detection zone, and produce detection signal, by the treatment and analysis of software, obtain the traffic parameters such as the volume of traffic, average speed, occupation rate, queue length.Computer vision technique can also be utilized to position vehicle, identify and follow the trail of, and the traffic behavior of detected object is analyzed and judges, finally complete the collection of various traffic flow data.Video-based vehicle detection is widely used in highway and urban road, is mainly used in the location of road conditions complexity at present, as interchange, ring road, tunnel, and the intersection etc. of urban road.Along with the improving constantly of Computer Vision and computer graphical recognition technology, the expansion of application and the reduction of hardware cost, the overall cost of video-based vehicle detection declines thereupon, and operation cost is lower the later stage in addition, and its range of application will constantly expand.But be used alone foreground detection techniques mostly in current video detection technology, make the accuracy detected be subject to the impact of circumstance complication degree very large, and the use of sorter largely solve this problem.

Summary of the invention

The invention provides a kind of video vehicle detection method merged based on multiple target tracking and cascade classifier, this method can realize vehicle detection accurately.

In order to solve the problem of vehicle detection accuracy, specific embodiment of the invention step is:

(1) sorter training, the multiple Haar-like feature through the screening of CART decision tree is trained;

(2) fresh target based on multi-scale filtering detects;

Further, the positive negative sample choosing proper ratio in step (1) carries out sorter training and refers to: by the positive sample of appropriate quantity and negative sample, from Sample Storehouse, random selecting is out, then 15 kinds of Haar-like features of positive negative sample are extracted, the feature being filtered out robust by CART decision tree carries out sorter training, obtains the cascade classifier of 20 grades.

Further, detecting based on the fresh target of multi-scale filtering in step (2) and refer to: moving object detection in video present frame is gone out, then by carrying out multi-scale filtering to targeted mass, adding the fresh target after filtering to tracker.

The invention has the advantages that and adopt the vehicle checking method that merges based on multiple target tracking and cascade classifier, the method adds sorter judgement tracing in present frame after all targets, thus is improve the degree of accuracy of detection.The algorithm complex of this method is lower in addition, can better adapt to the application of current computer vision system.

Accompanying drawing explanation

fig. 1for the implementing procedure of the embodiment of the present invention is illustrated figure;

fig. 2for positive sample instance figure;

fig. 3for negative sample example figure;

fig. 4for needing the feature mode extracted figure;

fig. 5for calculating examples of features figure;

fig. 6for the definition of RAST (x, y) figure

fig. 7for the positive and negative weights of the rectangle rotating 45 ° divide example figure

fig. 8for training classifier first order example figure;

fig. 9for testing result example figure;

Embodiment

In order to better the present invention is described, referring to accompanying drawingwith embodiment, further detailed description is done to specific embodiment of the invention.

as Fig. 1shown in, specific embodiment of the invention step is:

(1) train based on the multiple Haar-like feature of screening through CART decision tree.

First set up positive negative example base ( as Fig. 1,2shown in) description document, Postive.vec and Negtive.dat.In file, every a line stores the retrieving information of a pictures, is followed successively by image block title, image block at the starting position coordinates (left-top) of original image, the height and width of image block.

After description document has been set up, what carry out is image block haar feature extraction, feature mode as Fig. 4shown in, comprise altogether 15 kinds.Wherein the weights of black region are negative, and the weights of white portion are just.In order to improve counting yield, we use integral image to carry out the eigenwert of computed image block.Introduce the computation process of the rectangular characteristic value calculating vertical matrix type and rotate 45 ° below respectively.

The computation process of vertical eigenwert: fig. 5for the integral image of image, then the pixel value of a-quadrant (is designated as Pix _a), the pixel value in B region (is designated as Pix _b).

$\left.\begin{array}{c}{\mathrm{Pix}}_A={\mathrm{ii}}_8+{\mathrm{ii}}_4-{\mathrm{ii}}_5-{\mathrm{ii}}_7\\ {\mathrm{Pix}}_B={\mathrm{ii}}_9+{\mathrm{ii}}_5-{\mathrm{ii}}_6-{\mathrm{ii}}_8\\ feature={\mathrm{Pix}}_A-{\mathrm{Pix}}_B=\left({\mathrm{ii}}_8-{\mathrm{ii}}_7\right)+\left({\mathrm{ii}}_6-{\mathrm{ii}}_5\right)\\ -\left({\mathrm{ii}}_5-{\mathrm{ii}}_4\right)-\left({\mathrm{ii}}_9-{\mathrm{ii}}_8\right)\end{array}\right\}---\left(1\right)$

Rotate the computation process of the rectangular characteristic value of 45 °: fig. 6for the definition of RSAT (x, y) figure, fig. 7for the positive and negative weights of the rectangle rotating 45 ° divide example figure.

$\left.\begin{array}{c}RSAT\left(x,y\right)=\underset{x^{,}\≤x,x^{,}\≤x-\left|y-y^{,}\right|}{\Σ}I\left(x^{,},y^{,}\right)\\ RSAT\left(x,y\right)=RSAT\left(x-1,y-1\right)+RSAT\left(x-1,y\right)\\ +I\left(x,y\right)-RSAT\left(x-2,y-1\right)\end{array}\right\}---\left(2\right)$

Be input to extracting the feature having added label in sorter and train, training process optimum configurations: nstage=20, npos=1000, nneg=3000, w=40, h=40, other parameters are all suitable for default value.Finally train 20 grades of sorters that accuracy rate is 95.4%, fig. 8for first order training process.

(2) fresh target based on multi-scale filtering detects.

Foreground detection is mainly realized by the method for background modeling, and we use GMM background modeling here.It is that new agglomerate detects that agglomerate detects core: first from foreground image, detect all agglomerates, then overlapping agglomerate is had to abandon by less agglomerate (may be caused by noise) with tracked agglomerate, and remaining agglomerate is arranged according to size order, only retain wherein several larger agglomerate (being defaulted as 10).Finally utilize multi-scale filtering Rules Filtering, in screening, only have the targeted mass returning RECT result to be only standard compliant agglomerate, real new agglomerate is saved in agglomerate row in table.Now complete the task that fresh target detects, add fresh target to tracker, carry out subsequent treatment, final testing result as Fig. 9shown in.

The present embodiment adopts C++ programming realization in the computing machine being configured to 3.60GHzIntel (R) Xeon (R) E5-1620CPU and 8G internal memory, the image of process 21 frame 640*480 per second and Detection accuracy reaches 92.6%.

Claims (3)

1., based on the vehicle checking method that multiple target tracking and cascade classifier merge, it is characterized in that the concrete steps of the method are:

Step (1), sorter is trained, and the multiple Haar-like feature through the screening of CART decision tree is trained;

Step (2), the fresh target based on multi-scale filtering detects.

2. the vehicle checking method merged based on multiple target tracking and cascade classifier according to right 1, it is characterized in that sorter training in step (1), train based on the multiple Haar-like feature through the screening of CART decision tree, concrete steps are as follows:

S2.1: set up positive and negative pattern representation file.

S2.2: the 15 kinds of haar-like features extracting sample according to formula (1) (2), and carry out sorter training.

$\left.\begin{array}{c}{\mathrm{Pix}}_A={\mathrm{ii}}_8+{\mathrm{ii}}_4-{\mathrm{ii}}_5-{\mathrm{ii}}_7\\ {\mathrm{Pix}}_B={\mathrm{ii}}_9+{\mathrm{ii}}_5-{\mathrm{ii}}_6-{\mathrm{ii}}_8\\ feature={\mathrm{Pix}}_A-{\mathrm{Pix}}_B=({\mathrm{ii}}_8-{\mathrm{ii}}_7)+({\mathrm{ii}}_6-{\mathrm{ii}}_5)\\ -({\mathrm{ii}}_5-{\mathrm{ii}}_4)-({\mathrm{ii}}_9-{\mathrm{ii}}_8)\end{array}\right\}---\left(1\right)$

Pix in formula _aand Pix _bbe respectively the pixel value of region A and region B, ii _mrepresent the value on integral image block summit, feature is eigenwert.

The pixel that RSAT (x, y) is the 45 ° of regions in region, 45 °, point (x, the y) upper left corner and the lower left corner and, calculate the rectangular characteristic value of rotation 45 ° and calculate the difference being positioned at the capable rectangle RSAT (x, y) of cross exactly.

3. the vehicle checking method merged based on multiple target tracking and cascade classifier according to right 1, it is characterized in that the fresh target based on multi-scale filtering in step (3) detects, concrete steps are as follows:

S3.1: set up background model and carry out foreground detection.

S3.2: carry out new agglomerate detection according to the prospect detected.

S3.3: according to the foreground image in S3.1 and S3.2, carries out multiple dimensioned window filtering by prospect agglomerate, and window smallest dimension is that agglomerate is wide high by 1/2, and be agglomerate size to the maximum, the change of scale of window is 1.1.

S3.4: the agglomerate by filtering is demarcated position fresh target and adds tracker to.